RSNA PE Detection¶

Modality: CT | Format: DICOM (per-series volumes) | Dim: 3D | Labels: 13 PE labels

Overview¶

The RSNA-STR Pulmonary Embolism Detection Challenge (2020) dataset contains CT pulmonary angiography (CTPA) studies labelled for the presence and characteristics of pulmonary embolism. Each study has image-level (per-slice) and exam-level labels. The dataset contains approximately 7,279 training exams.

Download¶

Available on Kaggle: RSNA STR Pulmonary Embolism Detection. Requires Kaggle account.

Expected layout¶

rsna_pe_dataset/
  train.csv
  train/
    <StudyInstanceUID>/
      <SeriesInstanceUID>/
        <SOPInstanceUID>.dcm
        ...

Label columns¶

Column	Description
`acute_and_chronic_pe`	Acute and chronic PE
`central_pe`	Central PE
`chronic_pe`	Chronic PE only
`flow_artifact`	Flow artifact
`indeterminate`	Indeterminate
`leftsided_pe`	Left-sided PE
`negative_exam_for_pe`	No PE
`qa_contrast`	QA: contrast issue
`qa_motion`	QA: motion artifact
`rightsided_pe`	Right-sided PE
`rv_lv_ratio_gte_1`	RV/LV ratio ≥ 1
`rv_lv_ratio_lt_1`	RV/LV ratio < 1
`true_filling_defect_not_pe`	Filling defect, not PE

Constructor arguments¶

Argument	Type	Required	Default	Description
`base_image_dir`	`str`	Yes	`None`	`train/` (or `test/` for the test split) — direct parent of `<StudyInstanceUID>/<SeriesInstanceUID>/` (e.g. `/data/rsna_pe_dataset/train/`)
`csv_path`	`str`	No	auto	`train.csv`; auto-discovered
`hu_window`	`tuple[float,float]\\|None`	No	`(-1000, 1000)`	HU clip range for CT intensity normalisation

Shared arguments (inherited from `BaseRadiologicalDataset`)¶

Argument	Type	Required	Default	Description
`output_cls`	`bool`	No	`False`	Include `"cls"` in data dict
`output_mask`	`bool`	No	`False`	Not supported; silently ignored with a warning
`output_report`	`bool`	No	`False`	Not supported; silently ignored with a warning
`output_bbox`	`bool`	No	`False`	Not supported; silently ignored with a warning
`transform`	MONAI transform	No	`None`	MONAI Compose transform; `None` uses the default pipeline
`cache_dir`	`str`	No	`"./cache"`	MONAI cache directory
`dtype`	`torch.dtype`	No	`torch.bfloat16`	Output tensor dtype; use `torch.float32` on CPU
`harmonized_df`	`pd.DataFrame`	No	`None`	Pre-built harmonized DataFrame
`harmonizer`	harmonizer	No	`None`	Pre-instantiated harmonizer
`harmonizer_path`	`str`	No	`None`	Path to saved harmonized CSV

Dataset constructor¶

The constructor arguments above apply to both RSNAPEDetectionTrainDataset and RSNAPEDetectionTestDataset. The test split has no labels — output_cls is silently ignored.

Train split¶

from radharmony.dataset import RSNAPEDetectionTrainDataset

ds = RSNAPEDetectionTrainDataset(
    base_image_dir="/data/rsna_pe_dataset/train/",
    hu_window=(-1000, 1000),
    output_cls=True,
)
train_ds, val_ds = ds.get_datasets(n_splits=5)

Test split¶

from radharmony.dataset import RSNAPEDetectionTestDataset

ds = RSNAPEDetectionTestDataset(
    base_image_dir="/data/rsna_pe_dataset/test/",
    hu_window=(-1000, 1000),
)

Harmonizer¶

from radharmony.harmonizer import RSNAPEDetectionTrainHarmonizer

h = RSNAPEDetectionTrainHarmonizer(
    csv_path="/data/rsna_pe_dataset/train.csv",
    base_image_dir="/data/rsna_pe_dataset/train/",
)
df = h.harmonize()
print(df.columns.tolist())
df.to_csv("rsna_pe_harmonized.csv", index=False)

Load from saved harmonized CSV¶

import pandas as pd
from radharmony.dataset import RSNAPEDetectionTrainDataset

ds = RSNAPEDetectionTrainDataset(
    base_image_dir="/data/rsna_pe_dataset/train/",
    harmonized_df=pd.read_csv("rsna_pe_harmonized.csv"),
    output_cls=True,
)

Harmonizer notes¶

Each row in train.csv corresponds to one DICOM series (a full CT volume)
image_path is a relative path to the series directory, not an individual DICOM file
MONAI's ITKReader loads the entire series directory as a single 3D volume
Labels are exam-level (same label for all slices in a study)
HU window (-1000, 1000) is suitable for CTPA; adjust with hu_window= if needed

Outputs¶

Flag	Key	Shape	Notes
`output_cls=True`	`"cls"`	`(13,)`	Multi-label binary PE labels

Example paths¶

Role	Path
`base_image_dir` (train)	`/path/to/rsna_pe_dataset/train/`
`base_image_dir` (test)	`/path/to/rsna_pe_dataset/test/`
`csv_path` (`train.csv`)	`/path/to/rsna_pe_dataset/train.csv`
test CSV (`test.csv`)	`/path/to/rsna_pe_dataset/test.csv`